arxiv: 2605.03930 · v1 · submitted 2026-05-05 · 🪐 quant-ph

Recognition: unknown

Time-dependent variational Monte Carlo without bias

Wladislaw Krinitsin , Markus Schmitt

Authors on Pith no claims yet

Pith reviewed 2026-05-07 16:35 UTC · model grok-4.3

classification 🪐 quant-ph

keywords time-dependent variational Monte Carlounbiased estimationimportance samplingBorn distributionquench dynamicsneural quantum statesquantum many-body dynamics

0 comments

The pith

A cutoff-based deformation of the Born distribution enables unbiased time-dependent variational Monte Carlo for quantum dynamics.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes an unbiased version of time-dependent variational Monte Carlo that uses self-normalized importance sampling on a cutoff-deformed Born distribution. This addresses the estimation bias in traditional VMC, which can distort real-time evolution of quantum many-body systems. If successful, it allows more accurate simulations with expressive ansatze like neural quantum states during quench dynamics. The approach is tested in both pathological and generic cases, showing feasibility and accuracy.

Core claim

We propose an unbiased variant of time-dependent VMC using self-normalized importance sampling with respect to a cutoff-based deformation of the Born distribution. We demonstrate the feasibility and accuracy of the approach in pathological and generic cases of quench dynamics. Furthermore, we explore an alternative sampling strategy based on active learning via the tensor cross interpolation (TCI).

What carries the argument

self-normalized importance sampling with respect to a cutoff-based deformation of the Born distribution, which removes the estimation bias in time-dependent variational Monte Carlo

If this is right

Accurate modeling of real-time quantum many-body dynamics without systematic errors from sampling bias.
Compatibility with highly expressive ansatz functions such as neural quantum states.
Feasibility shown for both pathological and generic quench dynamics cases.
The TCI-based algorithm provides a complementary perspective to importance sampling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Such an unbiased method could enable reliable studies of long-time dynamics where bias accumulation might otherwise dominate.
Integration with tensor network methods might further improve efficiency in high-dimensional systems.
Testing on larger systems could reveal computational scaling advantages or limitations.

Load-bearing premise

The cutoff parameter in the deformed Born distribution can be selected to eliminate bias while keeping computational costs manageable and avoiding new uncontrolled errors across different quench scenarios.

What would settle it

A direct comparison in a small system with known exact dynamics, such as a spin chain quench, showing that the bias persists or that errors exceed those of standard methods would indicate the approach does not fully remove bias without new issues.

Figures

Figures reproduced from arXiv: 2605.03930 by Markus Schmitt, Wladislaw Krinitsin.

**Figure 1.** Figure 1: FIG. 1. MPS representation of view at source ↗

**Figure 2.** Figure 2: FIG. 2. MPS contractions between TCI view at source ↗

**Figure 4.** Figure 4: FIG. 4. Time evolution of an initially x-polarized state after view at source ↗

**Figure 5.** Figure 5: FIG. 5. (a) Time evolution of the ratio of normalization con view at source ↗

**Figure 6.** Figure 6: FIG. 6. Investigation of the relative error of the force vector view at source ↗

read the original abstract

When combined with highly expressive ansatz functions such as neural quantum states, variational Monte Carlo (VMC) constitutes a versatile numerical approach to tackle the quantum many-body problem in and out of equilibrium. However, its traditional formulation exhibits a subtle estimation bias leading to inaccuracies, which can be particularly detrimental when addressing real time dynamics. In this work, we investigate two avenues to circumvent said estimation bias. First, we propose an unbiased variant of time-dependent VMC using self-normalized importance sampling with respect to a cutoff-based deformation of the Born distribution. We demonstrate the feasibility and accuracy of the approach in pathological and generic cases of quench dynamics. Furthermore, we explore an alternative sampling strategy based on active learning via the tensor cross interpolation (TCI). While we find that our choice of tensor network architecture lacks the required low rank property, the proposed TCI-based algorithm complements the conventional importance sampling paradigm, providing an alternative perspective that may be further explored in future work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a concrete sampling fix to remove bias from real-time VMC and shows it works on the examples they ran, but the cutoff choice and variance behavior still need clearer rules.

read the letter

The main point is that they remove the known estimation bias in time-dependent VMC by switching to self-normalized importance sampling against a cutoff-deformed Born distribution. They test this on both pathological cases and ordinary quench dynamics and report that it stays accurate. That is the actual new piece: a direct modification of the sampling step that keeps the variational ansatz and the time-evolution equations unchanged while eliminating the bias term that usually appears in the estimator. The TCI route is presented as an exploratory alternative, but they openly note that their tensor-network architecture does not have the low-rank structure needed for it to be efficient, so that part functions more as a negative result than a competing method. Both moves are honest and stay within the existing VMC framework rather than inventing new variational principles. The soft spot is the cutoff itself. The abstract says a suitable deformation exists and can be chosen so that the importance weights correct the bias without new uncontrolled errors, yet it gives no explicit selection rule, no scaling of effective sample size with system size or time, and no bias-versus-cutoff curves. If the cutoff has to be tuned per system or if variance grows badly for longer evolutions, the method could still be practical only in limited regimes. The demonstrations are described as successful, but without the quantitative plots or implementation details it is hard to judge how generic the fix really is. This is for people already running neural-state VMC on dynamics problems and who have seen the bias issue in their own runs. A reader who needs a drop-in sampling change to reduce systematic error would find the proposal worth checking. The core construction is coherent enough and the authors are straightforward about the TCI limitation, so the paper deserves a serious referee rather than a desk rejection.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes an unbiased variant of time-dependent variational Monte Carlo (VMC) for quantum many-body systems, particularly when paired with neural quantum states. It addresses a subtle estimation bias in traditional TD-VMC by introducing self-normalized importance sampling with respect to a cutoff-based deformation of the Born distribution. Feasibility and accuracy are demonstrated for both pathological and generic cases of quench dynamics. The work also explores an alternative sampling strategy based on active learning via tensor cross interpolation (TCI), though the chosen tensor network architecture is found to lack the required low-rank property.

Significance. If the central claim holds, the unbiased TD-VMC approach would be a meaningful technical improvement for accurate real-time simulations of non-equilibrium quantum dynamics. Removing the estimation bias without prohibitive cost or new uncontrolled errors could enhance the reliability of VMC-based studies of quench dynamics and other time-dependent phenomena with expressive ansatze. The honest assessment of the TCI alternative's limitations is a positive aspect, as is the focus on a known but under-addressed bias issue in the field.

major comments (3)

[Abstract and proposed method section] The abstract and method description assert that self-normalized importance sampling w.r.t. the cutoff-deformed Born distribution yields an unbiased estimator, but no explicit derivation is provided showing that the deformation exactly corrects the original bias in the infinite-sample limit while preserving the time-dependent variational principle. This is load-bearing for the central claim of 'without bias'.
[Demonstrations in quench dynamics] No quantitative analysis (e.g., bias-vs-cutoff curves, effective sample size scaling with system size or evolution time, or variance bounds) is reported for the cutoff parameter choice in the generic quench cases. The skeptic concern that a generic cutoff may introduce uncontrolled errors or prohibitive cost is not addressed with concrete tests, undermining the feasibility demonstration.
[TCI exploration section] The TCI-based alternative is presented as complementary, but the finding that the tensor network architecture lacks the low-rank property is stated without details on the rank scaling or why TCI was still pursued; this weakens the claim that it provides a viable alternative perspective.

minor comments (2)

[Method] Notation for the deformed Born distribution and importance weights should be introduced with explicit equations early in the method section for clarity.
[Abstract] The abstract refers to 'pathological and generic cases' without defining what distinguishes them or referencing specific figures/tables; this should be clarified.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their careful reading of our manuscript and for providing constructive comments. We address each of the major comments below and outline the revisions we plan to make.

read point-by-point responses

Referee: [Abstract and proposed method section] The abstract and method description assert that self-normalized importance sampling w.r.t. the cutoff-deformed Born distribution yields an unbiased estimator, but no explicit derivation is provided showing that the deformation exactly corrects the original bias in the infinite-sample limit while preserving the time-dependent variational principle. This is load-bearing for the central claim of 'without bias'.

Authors: We agree that an explicit derivation is essential to substantiate the claim of unbiasedness. Although the method is motivated and described in the manuscript, we did not include a full step-by-step proof of unbiasedness in the infinite-sample limit. In the revised version, we will add a detailed derivation in the methods section demonstrating that the cutoff deformation, combined with self-normalized importance sampling, corrects the bias exactly as the sample size approaches infinity, while maintaining consistency with the time-dependent variational principle. revision: yes
Referee: [Demonstrations in quench dynamics] No quantitative analysis (e.g., bias-vs-cutoff curves, effective sample size scaling with system size or evolution time, or variance bounds) is reported for the cutoff parameter choice in the generic quench cases. The skeptic concern that a generic cutoff may introduce uncontrolled errors or prohibitive cost is not addressed with concrete tests, undermining the feasibility demonstration.

Authors: We acknowledge the value of quantitative analysis for the cutoff parameter. In the current manuscript, we demonstrate feasibility through accuracy in pathological and generic quench cases, but we did not provide systematic bias-vs-cutoff curves or scaling studies. To address this, we will include additional figures and analysis in the revised manuscript showing the dependence on the cutoff, effective sample sizes, and variance behavior for the generic cases, thereby mitigating concerns about uncontrolled errors or costs. revision: yes
Referee: [TCI exploration section] The TCI-based alternative is presented as complementary, but the finding that the tensor network architecture lacks the low-rank property is stated without details on the rank scaling or why TCI was still pursued; this weakens the claim that it provides a viable alternative perspective.

Authors: We appreciate this observation. The manuscript states the lack of low-rank property but provides limited details on the observed rank scaling or the rationale for including the TCI approach. In the revision, we will expand this section with specifics on the rank scaling behavior observed in our tensor network and clarify that TCI was explored to offer an alternative sampling perspective complementary to importance sampling, even if not optimal for the current architecture, as it may inspire future adaptations. revision: yes

Circularity Check

0 steps flagged

No circularity: method is a direct modification of existing VMC sampling with independent validation

full rationale

The paper proposes a new unbiased time-dependent VMC estimator via self-normalized importance sampling on a cutoff-deformed Born distribution, plus a TCI alternative. No load-bearing step reduces to a self-defined quantity, fitted parameter renamed as prediction, or self-citation chain. The central claim is supported by explicit demonstrations on quench dynamics (pathological and generic cases) rather than by construction from prior equations or author-specific uniqueness theorems. The derivation chain remains self-contained against external benchmarks, with the cutoff choice presented as a tunable but independently verifiable parameter rather than an implicit fit.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The approach introduces a cutoff parameter in the deformation of the Born distribution whose specific value must be chosen; no other free parameters, axioms, or invented entities are identifiable from the abstract.

free parameters (1)

cutoff parameter
Defines the deformation of the Born distribution used to enable unbiased sampling; its selection is central to the method's practicality and accuracy.

pith-pipeline@v0.9.0 · 5456 in / 1117 out tokens · 39616 ms · 2026-05-07T16:35:20.064053+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 9 canonical work pages

[1]

Carleo and M

G. Carleo and M. Troyer, Solving the quantum many- body problem with artificial neural networks, Science 355, 602 (2017)

2017
[2]

Sharir, A

O. Sharir, A. Shashua, and G. Carleo, Neural tensor con- tractions and the expressive power of deep neural quan- tum states, Physical Review B106, 205136 (2022)

2022
[3]

Hibat-Allah, M

M. Hibat-Allah, M. Ganahl, L. E. Hayward, R. G. Melko, and J. Carrasquilla, Recurrent neural network wave func- tions, Physical Review Research2, 023358 (2020)

2020
[4]

Schmitt and M

M. Schmitt and M. Heyl, Quantum many-body dynamics in two dimensions with artificial neural networks, Physi- cal Review Letters125, 100503 (2020)

2020
[5]

Burau and M

H. Burau and M. Heyl, Unitary long-time evolution with quantum renormalization groups and artificial neural networks, Physical Review Letters127, 050601 (2021)

2021
[6]

L. L. Viteritti, R. Rende, and F. Becca, Transformer vari- ational wave functions for frustrated quantum spin sys- tems, Physical Review Letters130, 236401 (2023)

2023
[7]

Donatella, Z

K. Donatella, Z. Denis, A. Le Boit´ e, and C. Ciuti, Dy- namics with autoregressive neural quantum states: Ap- plication to critical quench dynamics, Physical Review A 108, 022210 (2023)

2023
[8]

Chen and M

A. Chen and M. Heyl, Empowering deep neural quantum states through efficient optimization, Nature Physics20, 1476 (2024)

2024
[9]

A. Chen, V. D. Naik, and M. Heyl, Convolu- tional transformer wave functions, arXiv:2503.10462 10.48550/arXiv.2503.10462 (2025)

work page doi:10.48550/arxiv.2503.10462 2025
[10]

L. L. Viteritti, R. Rende, S. Sachdev, and G. Carleo, Ap- proaching the thermodynamic limit with neural-network quantum states (2026)

2026
[11]

Robledo Moreno, G

J. Robledo Moreno, G. Carleo, A. Georges, and J. Stokes, Fermionic wave functions from neural-network con- strained hidden states, Proceedings of the National Academy of Sciences119, e2122059119 (2022)

2022
[12]

J. Nys, G. Pescia, A. Sinibaldi, and G. Carleo, Ab-initio variational wave functions for the time-dependent many- electron schr¨ odinger equation, Nature communications 15, 9404 (2024)

2024
[13]

Y. Gu, W. Li, H. Lin, B. Zhan, R. Li, Y. Huang, D. He, Y. Wu, T. Xiang, M. Qin, L. Wang, and D. Lv, Solving the hubbard model with neural quantum states, arXiv:2507.02644 (2025)

work page arXiv 2025
[14]

Chen, Z.-Q

A. Chen, Z.-Q. Wan, A. Sengupta, A. Georges, and C. Roth, Neural network-augmented pfaffian wave- functions for scalable simulations of interacting fermions, arXiv:2507.10705 (2025)

work page arXiv 2025
[15]

Gauvin-Ndiaye, J

C. Gauvin-Ndiaye, J. Tindall, J. R. Moreno, and A. Georges, Mott transition and volume law entangle- ment with neural quantum states, Phys. Rev. Lett.134, 076502 (2025)

2025
[16]

L. L. Viteritti, R. Rende, C. Roth, A. Sengupta, G. Car- leo, and A. Georges, Beyond variational bias: Resolving intertwined orders in the hubbard model (2026)

2026
[17]

D. Luo, G. Carleo, B. K. Clark, and J. Stokes, Gauge equivariant neural networks for quantum lattice gauge theories, Phys. Rev. Lett.127, 276402 (2021)

2021
[18]

J. Lin, D. Luo, X. Yao, and P. E. Shanahan, Real-time dynamics of the schwinger model as an open quantum system with neural density operators, Journal of High Energy Physics2024, 1 (2024)

2024
[19]

M. J. Hartmann and G. Carleo, Neural-network approach to dissipative quantum many-body dynamics, Physical review letters122, 250502 (2019)

2019
[20]

M. Reh, M. Schmitt, and M. G¨ arttner, Time-dependent variational principle for open quantum systems with ar- tificial neural networks, Physical Review Letters127, 230501 (2021)

2021
[21]

D. Luo, Z. Chen, J. Carrasquilla, and B. K. Clark, Au- toregressive neural network for simulating open quantum systems via a probabilistic formulation, Physical review letters128, 090501 (2022)

2022
[22]

Mendes-Santos, M

T. Mendes-Santos, M. Schmitt, and M. Heyl, Highly re- solved spectral functions of two-dimensional systems with neural quantum states, Physical Review Letters131, 046501 (2023)

2023
[23]

Medvidovi´ c and D

M. Medvidovi´ c and D. Sels, Variational Quantum Dy- namics of Two-Dimensional Rotor Models, PRX Quan- tum4, 040302 (2023)

2023
[24]

Mendes-Santos, M

T. Mendes-Santos, M. Schmitt, A. Angelone, A. Ro- driguez, P. Scholl, H. Williams, D. Barredo, T. La- haye, A. Browaeys, M. Heyl, and M. Dalmonte, Wave- 10 Function Network Description and Kolmogorov Com- plexity of Quantum Many-Body Systems, Physical Re- view X14, 021029 (2024)

2024
[25]

Schmitt and M

M. Schmitt and M. Heyl, Simulating dynamics of correlated matter with neural quantum states, arXiv:2506.03124 (2025)

work page arXiv 2025
[26]

Van de Walle, M

A. Van de Walle, M. Schmitt, and A. Bohrdt, Many-body dynamics with explicitly time-dependent neural quan- tum states, Machine Learning: Science and Technology 6, 045011 (2025)

2025
[27]

Sinibaldi, D

A. Sinibaldi, D. Hendry, F. Vicentini, and G. Carleo, Time-dependent neural galerkin method for quantum dy- namics, Phys. Rev. Lett.136, 120402 (2026)

2026
[28]

Carleo, F

G. Carleo, F. Becca, M. Schir´ o, and M. Fabrizio, Local- ization and Glassy Dynamics Of Many-Body Quantum Systems, Scientific Reports2, 243 (2012)

2012
[29]

Sinibaldi, C

A. Sinibaldi, C. Giuliani, G. Carleo, and F. Vicentini, Un- biasing time-dependent variational monte carlo by pro- jected quantum evolution, Quantum7, 1131 (2023)

2023
[30]

F. J. R. Ruiz, M. K. Titsias, and D. M. Blei, Overdispersed black-box variational inference (2016), arXiv:1603.01140 [stat.ML]

work page arXiv 2016
[31]

K. Inui, Y. Kato, and Y. Motome, Determinant-free fermionic wave function using feed-forward neural net- works, Phys. Rev. Res.3, 043126 (2021)

2021
[32]

Misery, L

A. Misery, L. Gravina, A. Santini, and F. Vicentini, Look- ing elsewhere: improving variational monte carlo gradi- ents by importance sampling (2025), arXiv:2507.05352 [quant-ph]

work page arXiv 2025
[33]

Z.-Q. Wan, R. Wiersema, and S. Zhang, Removing nodal and support-mismatch pathologies in variational monte carlo via blurred sampling (2026), arXiv:2603.18148 [cond-mat.str-el]

work page arXiv 2026
[34]

Lao and S

Y. N´ u˜ nez Fern´ andez, M. K. Ritter, M. Jeannin, J.- W. Li, T. Kloss, T. Louvet, S. Terasaki, O. Parcollet, J. von Delft, H. Shinaoka, and X. Waintal, Learning ten- sor networks with tensor cross interpolation: New algo- rithms and libraries, SciPost Physics18, 10.21468/sci- postphys.18.3.104 (2025)

work page doi:10.21468/sci- 2025
[35]

Hackl, T

L. Hackl, T. Guaita, T. Shi, J. Haegeman, E. Demler, and J. I. Cirac, Geometry of variational methods: dynamics of closed quantum systems, SciPost Phys.9, 048 (2020)

2020
[36]

Hesterberg, Weighted average importance sampling and defensive mixture distributions, Technometrics37, 185 (1995)

T. Hesterberg, Weighted average importance sampling and defensive mixture distributions, Technometrics37, 185 (1995)

1995
[37]

A. B. Owen,Monte Carlo theory, methods and examples (https://artowen.su.domains/mc/, 2013)

2013
[38]

Casella and R

G. Casella and R. Berger,Statistical Inference(CRC Press, 2024)

2024
[39]

Gravina,Simulating quantum dynamics: neural- network quantum states, open quantum systems, and quantum devices, Ph.D

L. Gravina,Simulating quantum dynamics: neural- network quantum states, open quantum systems, and quantum devices, Ph.D. thesis, EPFL, Lausanne (2025)

2025
[40]

Carleo, Spectral and dynamical properties of strongly correlated systems (2011)

G. Carleo, Spectral and dynamical properties of strongly correlated systems (2011)

2011
[41]

D. Wu, L. Wang, and P. Zhang, Solving statistical me- chanics using variational autoregressive networks, Phys. Rev. Lett.122, 080602 (2019)

2019
[42]

Czischek, M

S. Czischek, M. G¨ arttner, and T. Gasenzer, Quenches near ising quantum criticality as a challenge for artificial neural networks, Physical Review B98, 024311 (2018)

2018
[43]

Liu and K

X. Liu and K. M. Goh, Resnet: Enabling deep convolu- tional neural networks through residual learning (2025), arXiv:2510.24036 [cs.CV]

work page arXiv 2025
[44]

We, however, resort to the plain stan- dard deviation, because vanishing SNR of zero compo- nents is not necessarily informative about the estimation cost

An alternative probe of fluctuations that is considered in recent work is the mean component-wise signal-to-noise- ratio (SNR) [32]. We, however, resort to the plain stan- dard deviation, because vanishing SNR of zero compo- nents is not necessarily informative about the estimation cost